Pith. sign in

REVIEW 3 major objections 6 minor 100 references

Learning to See Through Flare

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read NeuSee jointly learns a pupil-plane phase mask and a Mamba-GAN restorer, letting cameras image through laser irradiance up to one million times the sensor saturation threshold across the full visible spectrum.

desk verdict A credible simulation study of jointly learned laser-protection optics and restoration whose headline 10^6× suppression claim is unverified and internally inconsistent; worth peer review with major revisions. read the letter →

arxiv 2508.13907 v1 pith:2TPNATFQ submitted 2025-08-19 eess.IV cs.CV

Xiaopeng Peng , Heath Gemar , Erin Fleet , Kyle Novak , Abbie Watnik , Grover Swartzlander This is my paper · ORCID
classification eess.IVcs.CV
keywords learneddiffractiveopticalelementcomputationalimaginglaserdazzleprotectionsensorsaturationMamba-GANimagerestorationfull-spectrumphasemask
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces NeuSee, a computational imaging framework that pairs a learned diffractive optical element (a pupil-plane phase mask) with a frequency-space Mamba-GAN image-restoration network, trained end-to-end on 100K simulated images. Its central claim is that a camera equipped with this mask can keep imaging a scene while a laser at up to $10^6$ times the sensor saturation threshold $I_{\mathrm{sat}}$ strikes the sensor, across the full visible band (400-700 nm). The DOE spreads and suppresses the laser's peak irradiance before it reaches the sensor, and the restoration network inpaints residual saturation, removes blur and noise, and recovers the scene radiance. The authors report that NeuSee outperforms a heuristically learned half-ring DOE, improving restored-image quality by 10.1% (L1) and giving roughly 100 times stronger average suppression across visible laser wavelengths.

What carries the argument

The load-bearing object is the learned DOE height map $h_{\mathrm{DOE}}(u,v)$, generated by a UNet from pupil coordinates $(u,v)$, with phase $\phi_{\mathrm{DOE}}(u,v) = 2\pi\Delta n(\lambda) h_{\mathrm{DOE}}(u,v)/\lambda$. A Scaled Fresnel propagation turns this height map into wavelength-dependent point spread functions, so the mask is optimized against the full 31-band hyperspectral scene volume rather than three RGB channels. The laser is modeled as a plane wave that lands as a delta-function-like spot at the focal plane, while the background is convolved with the coded point spread function. Training uses two stages: Stage 1 jointly optimizes the mask and the restoration generator against multiscale discriminators with laser-suppression and background-transmission losses; Stage 2 freezes the DOE and fine-tunes the 8-layer FFT-Mamba restoration network with Charbonnier and Fourier-domain reconstruction losses.

What would settle it

Fabricate the learned DOE height map with the stated material dispersion, mount it in front of a camera whose parameters match Table 1, and measure the focal-plane point spread function and the peak irradiance from a 10 nm-bandwidth laser at $\alpha_l = 10^6 I_{\mathrm{sat}}$; if the measured suppression ratio or restored-image L1 diverges from the simulated values by more than the assumed noise levels, the central claim fails. A faster check is to measure the fabricated DOE's point spread function on an optical bench and compare it with the point spread function predicted by Eq. 4.

Watch

Extended reading notes

Core claim

NeuSee's central claim is that a single learned phase mask can simultaneously scatter an intense narrowband laser so its focal-plane peak falls below the damage threshold while ordinary scene light passes through, and that the residual saturated, blurred, noisy sensor image can be restored by a Mamba-GAN trained jointly with the mask. The paper claims suppression of peak laser irradiance up to $10^6$ times $I_{\mathrm{sat}}$, for laser wavelengths from 400 nm to 700 nm, under dynamically varying laser wavelength, intensity, position, ambient light, and sensor noise. In the reported simulations this full-spectrum protection comes with restored image quality 10.1% better on the L1 metric than the half-ring DOE baseline restored by the same network. The paper presents this as the first learned computational-imaging framework to achieve high-fidelity sensor protection across the whole visible spectrum.

Load-bearing premise

The physics-based simulator is faithful enough to a real fabricated DOE and camera that a mask trained purely in simulation will suppress a real laser at one million times saturation and let the restoration network recover the scene.

Editorial extensions

If this is right

  • If the central claim holds, a single passive pupil-plane mask can replace wavelength-specific optical limiters for visible-band laser protection, since the same DOE covers 400-700 nm.
  • A camera using NeuSee would keep functioning at laser irradiances up to $10^6 I_{\mathrm{sat}}$, the regime where silicon sensors begin to risk permanent damage, so the protection is not just dazzle reduction.
  • Because the only latency is the restoration network's post-processing, the optical mask itself provides instantaneous, linear, broadband protection without moving parts or power.
  • The 10.1% quality gain over the half-ring baseline indicates that jointly learned masks encode scene information more efficiently than heuristic mask shapes, suggesting further gains from richer mask parameterizations.
  • Deployed on autonomous vehicles, robots, security cameras, or augmented-reality headsets, the system would preserve vision during deliberate laser attacks or accidental laser exposure rather than blinding the platform.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The obvious next test is fabrication: etching the learned height map and measuring the on-bench point spread function and laser suppression ratio, since the paper contains no real experiment or measured point spread function.
  • The framework should extend to simultaneous multi-wavelength or out-of-band lasers by retraining with additional spectral bands, because the architecture is wavelength-agnostic apart from the material dispersion model.
  • The same joint mask-plus-restoration recipe could be applied to other saturation sources, such as sun glare or high-dynamic-range clipping, where the mask spreads the energy and the network inpaints the clipped region.
  • The 100 times average suppression advantage over the half-ring mask is a simulated quantity; if the simulator's delta-function laser model overstates the focused peak, real-world suppression could be lower, so the $10^6$ figure should be read as simulation-bound until hardware validation.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces NeuSee, an end-to-end learned computational imaging framework that jointly optimizes a diffractive optical element (DOE) represented by a UNet and a frequency-space Mamba-GAN image restoration network. The system is trained on 100K simulated RGB images converted to 31-band hyperspectral radiance, with a physics-based forward model that adds laser dazzle, lens flare, sensor noise, and saturation. The authors claim that NeuSee suppresses peak laser irradiance up to 10^6 times the sensor saturation threshold, works across the full visible spectrum (400–700 nm), and improves restored image quality by 10.1% over a half-ring DOE baseline.

Significance. If the central claims are verified, the work would be a meaningful advance in computational imaging for laser protection, combining learned diffractive optics with learned restoration and addressing a realistic threat model (dynamic laser wavelength, intensity, position, and ambient conditions). The paper ships a detailed simulation pipeline, a sizable 100K-image training regimen, and a two-stage training strategy to balance conflicting DOE and restoration objectives. Strengths include the explicit treatment of hyperspectral propagation, sensor noise statistics, and the use of adversarial loss for restoring saturated regions. However, the headline quantitative claim — suppression of peak laser irradiance up to 10^6× I_sat — is not directly evidenced anywhere in the manuscript, and the evaluation is entirely simulation-based with no hardware verification. The significance is therefore conditional on the authors supplying the missing absolute suppression metric and reconciling internal inconsistencies.

major comments (3)
  1. [Section 6, Eq. (7b) and Eq. (13)] The central claim of suppressing peak laser irradiance up to 10^6× I_sat is not supported by any reported absolute LSR value. From Eq. (7b), Il,peak = αl·LSR·Isat, so for αl = 10^6 the required LSR is at most 10^-6. Although Section 6 states that NeuSee achieves '5× stronger suppression' (text) and '100 times stronger suppression' (Fig. 6 caption) than the half-ring mask, no absolute LSR value or its wavelength dependence is reported. Moreover, the DOE objective in Eq. (13), LDOE(LSR)=Σ Il(λ)/Il0(λ), is an integrated on-sensor energy ratio, not a peak-irradiance ratio; minimizing this sum can be satisfied by redistributing energy over many pixels even when the peak LSR remains far above 10^-6. The paper must report the actual peak LSR as a function of wavelength, clarify whether the 10^6 figure refers to the training range [0, 2e6] in Section 5 or to achieved suppression, and reconcile the 5×/100× discrepancy.
  2. [Sections 3.4 and 7] The paper's claims are presented in the abstract and introduction as achieved sensor-protection capabilities, but the entire evaluation is simulated. Section 3.4 states that simulation parameter values 'match the experiment', yet no experimental setup, fabricated DOE, measured PSF, or real camera image appears in the paper. The conclusion appropriately says 'In simulation', but the abstract and introduction do not carry this qualifier. Because the learned DOE and restoration network were trained entirely in simulation, the transfer of these results to real hardware is an untested assumption. The authors should either add hardware validation or consistently qualify all headline claims as simulation-based.
  3. [Section 6, 'outperforms other learned DOEs'] The claim that NeuSee 'outperforms other learned DOEs' is supported only by a comparison to a single half-ring mask trained with a heuristic method [30]. The paper does not compare against other learned DOEs, such as those in Refs. [31] and [33], nor against a restoration-only baseline without any DOE. The quantitative 10.1% L1 improvement is reported as a single average over 7K test images without error bars, per-condition breakdown, or statistical significance testing. Given the strong comparative wording in the contributions list, the evaluation should include at least one additional learned-DOE baseline and report variability over test conditions.
minor comments (6)
  1. [Abstract] The phrase 'suppress the peak laser irradiance as high as 10^6 times the sensor saturation threshold' is ambiguous: it could mean the system handles incident lasers up to 10^6× I_sat or that it achieves an LSR of 10^-6. Please rephrase to state the intended meaning explicitly.
  2. [Section 3.3, Eq. (5b)] The laser is modeled as a Dirac delta function with no spatial extent; real laser beams have finite spot size and angular divergence. This simplification should be stated as a limitation, and its effect on the reported suppression ratios should be discussed.
  3. [Section 5, paragraph beginning 'Deep learning systems'] The description 'Laser strengths αl are randomly sampled from 100K predetermined values, which are uniformly distributed in the range [0, 2e6]' is unclear: are the same 100K values reused across training iterations, or independently re-sampled each iteration? Please clarify.
  4. [Section 6, Fig. 6 caption] The caption states '100 times stronger suppression' while the running text states '5× stronger suppression'. These should be reconciled, ideally with the numerical values of LSR for both systems included in the figure or table.
  5. [Section 4, Eq. (12b) and Eq. (14)] There are formatting errors: Eq. (12b) contains 'DL 2 (bL)' where the square is misplaced, and Eq. (14) has an unbalanced parenthesis in the FFT objective ('|F(bL)−F (ˆbL)|'). Correct these for clarity.
  6. [References] Reference [92] is a GitHub URL without version, commit hash, or date of access; the 'Scaled Fresnel method' should be cited to a peer-reviewed publication with a precise algorithm description.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the physics-based simulation and learned DOE/restoration pipeline are self-contained; the 10^6 suppression claim is a training-range and evaluation condition, not a definitional equivalent of the model output.

full rationale

I find no circular step in the paper's derivation chain. The DOE height is produced by a neural representation net from pupil coordinates (Eq. 2), mapped to phase through Eq. 3; the focal-plane field and PSF follow from scalar diffraction (Eqs. 4a-4b); the sensor irradiance is modeled in Eqs. 5a-5b; and restoration is a learned map from the sensor image to scene radiance (Eq. 11). Each stage is an input assumption or an optimization target, not an output that is fed back to define the input. The headline claim of suppressing peak irradiance up to 10^6 times saturation is best read as a statement about the training and evaluation range: Section 5 states that laser strengths are sampled uniformly in [0,2e6], and Section 6 evaluates at alpha_l up to 1e6. The DOE loss in Eq. 13 does include LSR as a minimization objective, but reporting performance on a held-out simulated test set after optimizing a loss is standard supervised learning practice; the mask could in principle fail to suppress, so the result is not true by construction. The main weaknesses are verification gaps, not circularity: no absolute LSR value is reported, the text (5x) and Fig. 6 caption (100x) disagree on the relative improvement, and the physical simulation parameters are claimed to match an experiment that is not shown. These are correctness and reporting concerns. The half-ring baseline comes from the authors' prior works [28,30], but it is used as a comparison baseline rather than as load-bearing justification for the physics or the method, so it does not constitute circular self-citation. No uniqueness theorem, hidden ansatz, or renaming of a known result is invoked to force the conclusions.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the simulator being a faithful model of a real DOE plus camera: the propagation model, the plane-wave laser model, the noise model, and the MST++ spectral reconstruction are all unvalidated assumptions. There are no invented physical entities. The '10^6 suppression' headline is bounded by the chosen training range and by loss weights that are set by hand.

free parameters (4)
  • Laser strength training range [0, 2e6] times I_sat = [0, 2e6]
    Uniformly sampled in Section 5; the claimed 10^6 handling is this chosen range, not a measured physical limit.
  • Laser spectral bandwidth Delta_lambda_FWHM = 10 nm = 10 nm
    Gaussian profile in Section 3.4; chosen by hand for the simulation, not tied to a specific laser.
  • GAN loss weights lambda_ADV=0.1, lambda_GP=1 = 0.1, 1.0
    Hyperparameters in Eq. 12; chosen by hand, not fitted, but they shape the restoration quality.
  • DOE and restoration loss weighting (no scaling between LDOE and GAN terms) = unstated
    Equation 15 sums LDOE and LGAN without explicit weights; the balance is implicitly chosen by the optimization, and no sensitivity analysis is provided.
assumptions (6)
  • domain assumption Simulator parameters in Table 1 match the real experimental camera, and the simulation-to-real transfer is lossless.
    The paper states the parameters 'match the experiment' but presents no real experiment; this is the load-bearing premise for all claimed capabilities.
  • domain assumption Scaled Fresnel propagation (Eq. 4) accurately models the PSF of the pupil-masked system.
    Used throughout Section 3; correctness of the simulated PSF is the basis for all laser suppression and image formation results.
  • domain assumption Shift-invariant imaging with spatial convolution (Eq. 5a) is valid for the large field-of-view scenes.
    The restoration network relies on this model for the background; real lenses may have field-dependent aberrations.
  • domain assumption The laser is a plane wave at the entrance pupil and a delta function in the focal plane (Eq. 5b).
    Real laser beams have finite size, coherence, and divergence; the paper samples incident angles but not beam shape.
  • domain assumption MST++ (Eq. 1) reconstructs a sufficiently accurate 31-band HSI from RGB for DOE training.
    The full visible spectrum claim rests on this pretrained black box; no HSI ground truth is used.
  • domain assumption The DOE height map (Eq. 2) can be manufactured with the corresponding phase profile and dispersion Delta_n(lambda).
    No fabrication or measured height/phase is reported; manufacturability is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to See Through Flare." pith.science (2026). https://pith.science/paper/2TPNATFQ

@misc{pith2026250813907,
  author       = {Pith},
  title        = {Pith review of: Learning to See Through Flare},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2TPNATFQ}},
  note         = {Machine review of arXiv:2508.13907}
}
abstract

Machine vision systems are susceptible to laser flare, where unwanted intense laser illumination blinds and distorts its perception of the environment through oversaturation or permanent damage to sensor pixels. We introduce NeuSee, the first computational imaging framework for high-fidelity sensor protection across the full visible spectrum. It jointly learns a neural representation of a diffractive optical element (DOE) and a frequency-space Mamba-GAN network for image restoration. NeuSee system is adversarially trained end-to-end on 100K unique images to suppress the peak laser irradiance as high as $10^6$ times the sensor saturation threshold $I_{\textrm{sat}}$, the point at which camera sensors may experience damage without the DOE. Our system leverages heterogeneous data and model parallelism for distributed computing, integrating hyperspectral information and multiple neural networks for realistic simulation and image restoration. NeuSee takes into account open-world scenes with dynamically varying laser wavelengths, intensities, and positions, as well as lens flare effects, unknown ambient lighting conditions, and sensor noises. It outperforms other learned DOEs, achieving full-spectrum imaging and laser suppression for the first time, with a 10.1\% improvement in restored image quality.

Figures

Figures reproduced from arXiv: 2508.13907 by the authors.

Figure 1
Figure 1. Illustration of sensor damage risks under the laser illumi [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The proposed NeuSee system is a jointly learned imaging framework designed to protect sensors from laser dazzle across the full [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Spectral profile of scene illumination and a red laser. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Lens flare effects under laser illumination at various [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Illustration of heterogeneous data and model parallelism for distributed end-to-end training of the NeuSee system. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Comparison of laser suppression ratio, DOE (or phase [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Row 1 compares sensor image of NeuSee (learned in stage-1), the heuristically learned half-ring DOE [ [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

100 extracted references · 70 canonical work pages

  1. [30]

    Learning to See Through Dazzle

    Xiaopeng Peng, Erin F Fleet, Abbie T Watnik, and Grover A Swartzlander. Learning to see through dazzle.arXiv preprint arXiv:2402.15919, 2024. 7, 8

    work page Pith review arXiv 2024

  2. [31]

    Learned phase mask to protect cam- era under laser irradiation

    Junyu Zhang, Qing Ye, Yunlong Wu, Yangliang Li, Yihua Hu, and Haoqi Luo. Learned phase mask to protect cam- era under laser irradiation. Optics Express, 32(24):42674– 42691, 2024

    work page 2024

  3. [33]

    Laser protection via jointly learned defocus and image reconstruction

    Johannes Meyer, Michael Henrichsen, Christian Eisele, Bas- tian Schwarz, J¨urgen Limbach, Gunnar Ritt, Stefanie Den- gler, Lukas Dippon, and Christian Kludt. Laser protection via jointly learned defocus and image reconstruction. Au- thorea Preprints, 2025. 1

    work page 2025

  4. [1]

    The potential role of laser in combating uav: part 2; laser as a countermeasure and weapon

    Ove Steinvall. The potential role of laser in combating uav: part 2; laser as a countermeasure and weapon. In Technologies for Optical Countermeasures XVIII and High- Power Lasers: Technology and Systems, Platforms, Effects V, volume 11867. SPIE, 2021. 1

    2021

  5. [2]

    Laser dazzling: an overview

    Ove Steinvall. Laser dazzling: an overview. In Technologies for Optical Countermeasures XIX , volume 12738, pages 17–31. SPIE, 2023

    2023

  6. [3]

    The disruptive impact of dynamic laser dazzling on template matching algorithms applied to thermal infrared imagery

    Gareth D Lewis, Alexander Borghgraef, and Marijke Van- dewal. The disruptive impact of dynamic laser dazzling on template matching algorithms applied to thermal infrared imagery. In Technologies for Optical Countermeasures XIX, volume 12738, page 1273803. SPIE, 2023. 1

    2023

  7. [4]

    Adversarial laser beam: Effective physical-world attack to dnns in a blink

    Ranjie Duan, Xiaofeng Mao, A Kai Qin, Yuefeng Chen, Shaokai Ye, Yuan He, and Yun Yang. Adversarial laser beam: Effective physical-world attack to dnns in a blink. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16062–16071, 2021. 1

    2021

  8. [5]

    Engineering pupil func- tion for optical adversarial attacks

    Kyulim Kim, Jeongsoo Kim, Seungri Song, Jun-Ho Choi, Chulmin Joo, and Jong-Seok Lee. Engineering pupil func- tion for optical adversarial attacks. Optics Express, 30(5): 6500–6518, 2022

    2022

Show all 100 references
  1. [6]

    Embodied laser attack: Leveraging scene priors to achieve agent-based robust non-contact attacks

    Yitong Sun, Yao Huang, and Xingxing Wei. Embodied laser attack: Leveraging scene priors to achieve agent-based robust non-contact attacks. In ACM Multimedia 2024. 1

    2024

  2. [7]

    Beyond laser safety glasses: augmented reality in optics laboratories

    Franco Quercioli. Beyond laser safety glasses: augmented reality in optics laboratories. Applied optics, 56(4):1148– 1150, 2017. 1

    2017

  3. [8]

    Virtual reality (vr) for laser safety training

    Grzegorz Owczarek, Mieszko Wodzy ´nski, Joanna Szkud- larek, and Marcin Jachowicz. Virtual reality (vr) for laser safety training. In 2021 IEEE 2nd International Conference on Human-Machine Systems (ICHMS) , pages 1–3. IEEE, 2021

    2021

  4. [9]

    Occupational eye protection us- ing augmented reality: a proof of concept

    JM Deniel and S Thommet. Occupational eye protection us- ing augmented reality: a proof of concept. Radioprotection, 57(2):165–173, 2022

    2022

  5. [10]

    Mixed reality for laser safety at advanced op- tics laboratories

    Ke Li, Aradhana Choudhuri, Susanne Schmidt, Tino Lang, Reinhard Bacher, Ingmar Hartl, Wim Leemans, and Frank Steinicke. Mixed reality for laser safety at advanced op- tics laboratories. In International Laser Safety Conference, number PUBDB-2023-07345. Control-System, 2023. 1

    2023

  6. [11]

    Laser incidents. www . faa . gov / about / initiatives/lasers/laws. 1

  7. [12]

    www.osha.gov/laser-hazards

    Laser hazard. www.osha.gov/laser-hazards. 1

  8. [13]

    Laser safety eyewear

    Deepthi Malayanur, Venkataram Nagaraj Mysore, et al. Laser safety eyewear. CosmoDerma, 2, 2022. 1

    2022

  9. [14]

    www.ilda.com/camera-sensor-damage.htm. 1

  10. [15]

    Laser safety calculations for imaging sensors

    Gunnar Ritt. Laser safety calculations for imaging sensors. Sensors, 19(17), 2019. 1

    2019

  11. [16]

    Damage thresholds of silicon-based cameras for in-band and out-of-band laser expositions

    Francis Th ´eberge, Michel Auclair, Jean-Fran c ¸ois Daigle, and Dominik Pudo. Damage thresholds of silicon-based cameras for in-band and out-of-band laser expositions. Ap- plied Optics, 61(10):2473–2482, 2022. 1

    2022

  12. [17]

    Preventing image information loss of imaging sensors in case of laser dazzle

    Gunnar Ritt, Bastian Schwarz, and Bernd Eberle. Preventing image information loss of imaging sensors in case of laser dazzle. Optical Engineering, 58(1):013109, 2019. 1

    2019

  13. [18]

    Use of complementary wave- length bands for laser dazzle protection

    Gunnar Ritt and Bernd Eberle. Use of complementary wave- length bands for laser dazzle protection. Optical Engineer- ing, 59(1):015106, 2020. 1

    2020

  14. [19]

    Thermoset polymers as host for opti- cal limiting

    Jade Caillieaudeaux, Olivier Muller, Morgane Guerchoux, C´elia Bruder, Lionel Merlat, Anne-Sophie Schuller, and Christelle Delaite. Thermoset polymers as host for opti- cal limiting. Journal of Applied Polymer Science, 141(3): e54810, 2024. 1

    2024

  15. [20]

    Self-activating liquid crystal devices for smart laser protection

    Ling Wang. Self-activating liquid crystal devices for smart laser protection. Liquid Crystals, 43(13-15):2062–2078,

  16. [21]

    Ad- vanced liquid crystal-based switchable optical devices for light protection applications: principles and strategies

    Ruicong Zhang, Zhibo Zhang, Jiecai Han, Lei Yang, Jia- jun Li, Zicheng Song, Tianyu Wang, and Jiaqi Zhu. Ad- vanced liquid crystal-based switchable optical devices for light protection applications: principles and strategies. Light: Science & Applications, 12(1):11, 2023. 1

    2023

  17. [22]

    Optical limiting based on huygens’ metasurfaces

    Austin Howes, Zhihua Zhu, David Curie, Jason R Avila, Virginia D Wheeler, Richard F Haglund, and Jason G Valen- tine. Optical limiting based on huygens’ metasurfaces. Nano Letters, 20(6):4638–4644, 2020. 1

    2020

  18. [23]

    Linear-to-circular po- larization conversion with full-silica meta-optics to reduce nonlinear effects in high-energy lasers

    Nicolas Bonod, Pierre Brianceau, J´erˆome Daurios, Sylvain Grosjean, Nadja Roquin, Jean-Francois Gleyze, Laurent Lamaign`ere, and J´erˆome Neauport. Linear-to-circular po- larization conversion with full-silica meta-optics to reduce nonlinear effects in high-energy lasers. Nat...

    2023

  19. [24]

    Mitigation of laser dazzle effects on a mid-wave infrared thermal imager by reducing the integration time of the focal plane array

    GD Lewis, CN Santos, and M Vandewal. Mitigation of laser dazzle effects on a mid-wave infrared thermal imager by reducing the integration time of the focal plane array. In Technologies for Optical Countermeasures XVI, volume 11161, page 1116108. International Society for Optic...

    2019

  20. [25]

    Smoke as protection against high energy laser effects

    Ric HMA Schleijpen, Sven Binsbergen, Amir V osteen, Karin de Groot-Trouw, Denise Meuken, and Alexander MJ Van Eijk. Smoke as protection against high energy laser effects. In Technologies for Optical Countermeasures XVIII and High-Power Lasers: Technology and Systems, Plat- for...

    2021

  21. [26]

    Reducing the risk of laser damage in a focal plane array using linear pupil-plane phase elements

    Garreth J Ruane, Abbie T Watnik, and Grover A Swartzlan- der. Reducing the risk of laser damage in a focal plane array using linear pupil-plane phase elements. Applied optics, 54 (2):210–218, 2015. 1

    2015

  22. [27]

    Performance analysis of spiral axi- con wavefront coding imaging system for laser protection

    Haoqi Luo, Yangliang Li, Junyu Zhang, Hao Zhang, Yun- long Wu, and Qing Ye. Performance analysis of spiral axi- con wavefront coding imaging system for laser protection. Current Optics and Photonics, 8(4):355–365, 2024. 1

    2024

  23. [28]

    Half-ring point spread functions

    Jacob H Wirth, Abbie T Watnik, and Grover A Swartzlander. Half-ring point spread functions. Optics letters, 45(8):2179– 2182, 2020. 1, 7

    2020

  24. [29]

    Computational Imaging and Its Applications

    Helen Peng. Computational Imaging and Its Applications. Rochester Institute of Technology, 2022. 2

    2022

  25. [32]

    Opsecurecam: optically enhanced secure camera via an engineering point spread function

    Haoqi Luo, Junyu Zhang, Ye Liu, Weibing Sun, Yunlong Wu, Qing Ye, and Yihua Hu. Opsecurecam: optically enhanced secure camera via an engineering point spread function. Op- tics Express, 33(10):20880–20893, 2025

    2025

  26. [34]

    Coded exposure photography: motion deblurring using fluttered shutter

    Ramesh Raskar, Amit Agrawal, and Jack Tumblin. Coded exposure photography: motion deblurring using fluttered shutter. In Acm Siggraph 2006 Papers , pages 795–804

    2006

  27. [35]

    The diffractive achromat full spectrum computational imag- ing with diffractive optics

    Yifan Peng, Qiang Fu, Felix Heide, and Wolfgang Heidrich. The diffractive achromat full spectrum computational imag- ing with diffractive optics. ACM Transactions on Graphics (TOG), 35(4):1–11, 2016. 2

    2016

  28. [36]

    Learned rotationally symmetric diffractive achromat for full-spectrum computa- tional imaging

    Xiong Dun, Hayato Ikoma, Gordon Wetzstein, Zhanshan Wang, Xinbin Cheng, and Yifan Peng. Learned rotationally symmetric diffractive achromat for full-spectrum computa- tional imaging. Optica, 7(8):913–922, 2020. 2

    2020

  29. [37]

    Learning rank-1 diffractive optics for single- shot high dynamic range imaging

    Qilin Sun, Ethan Tseng, Qiang Fu, Wolfgang Heidrich, and Felix Heide. Learning rank-1 diffractive optics for single- shot high dynamic range imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition, pages 1386–1396, 2020. 2

    2020

  30. [38]

    Deep optics for single-shot high-dynamic- range imaging

    Christopher A Metzler, Hayato Ikoma, Yifan Peng, and Gor- don Wetzstein. Deep optics for single-shot high-dynamic- range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1375– 1385, 2020. 2

    2020

  31. [39]

    Glare aware photography: 4d ray sampling for reducing glare effects of camera lenses

    Ramesh Raskar, Amit Agrawal, Cyrus A Wilson, and Ashok Veeraraghavan. Glare aware photography: 4d ray sampling for reducing glare effects of camera lenses. In ACM SIG- GRAPH 2008 papers, pages 1–10. 2008. 2

    2008

  32. [40]

    Glare encoding of high dynamic range images

    Mushfiqur Rouf, Rafał Mantiuk, Wolfgang Heidrich, Matthew Trentacoste, and Cheryl Lau. Glare encoding of high dynamic range images. In CVPR 2011, pages 289–296. IEEE, 2011. 2

    2011

  33. [41]

    End-to-end optimization of optics and image processing for achromatic extended depth of field and super- resolution imaging

    Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and super- resolution imaging. ACM Transactions on Graphic...

    2018

  34. [42]

    Phase- cam3d—learning phase masks for passive single view depth estimation

    Yicheng Wu, Vivek Boominathan, Huaijin Chen, Aswin Sankaranarayanan, and Ashok Veeraraghavan. Phase- cam3d—learning phase masks for passive single view depth estimation. In 2019 IEEE International Conference on Com- putational Photography (ICCP), pages 1–12. IEEE, 2019

    2019

  35. [43]

    Codedstereo: Learned phase masks for large depth-of- field stereo

    Shiyu Tan, Yicheng Wu, Shoou-I Yu, and Ashok Veeraragha- van. Codedstereo: Learned phase masks for large depth-of- field stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7170–7179,

  36. [44]

    Dappled photography: Mask enhanced cameras for heterodyned light fields and coded aperture refocusing

    Ashok Veeraraghavan, Ramesh Raskar, Amit Agrawal, Ankit Mohan, and Jack Tumblin. Dappled photography: Mask enhanced cameras for heterodyned light fields and coded aperture refocusing. ACM Transactions on Graphics (TOG), 26(3):69, 2007. 2

    2007

  37. [45]

    Shape recon- struction and orientation estimation of transparent micro- scopic object using light field microscopy

    Xiaopeng Peng and Grover A Swartzlander. Shape recon- struction and orientation estimation of transparent micro- scopic object using light field microscopy. In Frontiers in Optics, pages FTh5C–3. Optica Publishing Group, 2016. 2

    2016

  38. [46]

    Recent advances in lensless imaging

    Vivek Boominathan, Jacob T Robinson, Laura Waller, and Ashok Veeraraghavan. Recent advances in lensless imaging. Optica, 9(1):1–16, 2022. 2

    2022

  39. [47]

    Learning privacy-preserving optics for human pose estima- tion

    Carlos Hinojosa, Juan Carlos Niebles, and Henry Arguello. Learning privacy-preserving optics for human pose estima- tion. In Proceedings of the IEEE/CVF international confer- ence on computer vision, pages 2573–2582, 2021. 2

    2021

  40. [48]

    Learning phase mask for privacy- preserving passive depth estimation

    Zaid Tasneem, Giovanni Milione, Yi-Hsuan Tsai, Xiang Yu, Ashok Veeraraghavan, Manmohan Chandraker, and Francesco Pittaluga. Learning phase mask for privacy- preserving passive depth estimation. In Computer Vision– ECCV 2022: 17th European Conference, Tel Aviv, Israel, October ...

    2022

  41. [49]

    Randomized aperture imaging

    Xiaopeng Peng, Garreth J Ruane, and Grover A Swartz- lander Jr. Randomized aperture imaging. arXiv preprint arXiv:1601.00033, 2016. 2

    2016 arXiv

  42. [50]

    Mirror swarm space telescope

    Xiaopeng Peng and Grover A Swartzlander. Mirror swarm space telescope. In 2014 IEEE Western New York Image and Signal Processing Workshop (WNYISPW), pages 37–41. IEEE, 2014. 2

    2014

  43. [51]

    Com- pact snapshot hyperspectral imaging with diffracted rotation

    Daniel S Jeon, Seung-Hwan Baek, Shinyoung Yi, Qiang Fu, Xiong Dun, Wolfgang Heidrich, and Min H Kim. Com- pact snapshot hyperspectral imaging with diffracted rotation. ACM Transactions on Graphics (TOG), 38(4):1–13, 2019. 2

    2019

  44. [52]

    Single-shot hyperspectral-depth imaging with learned diffractive optics

    Seung-Hwan Baek, Hayato Ikoma, Daniel S Jeon, Yuqi Li, Wolfgang Heidrich, Gordon Wetzstein, and Min H Kim. Single-shot hyperspectral-depth imaging with learned diffractive optics. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 2651–2660,

  45. [53]

    P-48: Evalua- tion of field of view in optical see-through near eye displays

    Xi Mou, Xiaopeng Peng, and Jianping Wang. P-48: Evalua- tion of field of view in optical see-through near eye displays. In SID Symposium Digest of Technical Papers, volume 55, pages 1548–1550. Wiley Online Library, 2024. 2

    2024

  46. [54]

    38-2: Invited paper: Evaluating optical performance and image quality in augmented reality eyewear: Standardization, challenges, and measurement methods

    Xi Mou, Xiaopeng Peng, and Tongsheng Mou. 38-2: Invited paper: Evaluating optical performance and image quality in augmented reality eyewear: Standardization, challenges, and measurement methods. In SID Symposium Digest of Technical Papers, volume 55, pages 324–326. Wiley Onli...

    2024

  47. [55]

    Rf- sauron: Enabling contact-free interaction on eyeglass using conformal rfid tag

    Baizhou Yang, Ling Chen, Xiaopeng Peng, Jiashen Chen, Yani Tang, Wei Wang, Dingyi Fang, and Chao Feng. Rf- sauron: Enabling contact-free interaction on eyeglass using conformal rfid tag. IEEE Internet of Things Journal, 2025. 2

    2025

  48. [56]

    End-to-end learned, optically coded super-resolution spad camera

    Qilin Sun, Jian Zhang, Xiong Dun, Bernard Ghanem, Yifan Peng, and Wolfgang Heidrich. End-to-end learned, optically coded super-resolution spad camera. ACM Transactions on Graphics (TOG), 39(2):1–14, 2020. 2

    2020

  49. [57]

    Seeing through obstructions with diffractive cloaking

    Zheng Shi, Yuval Bahat, Seung-Hwan Baek, Qiang Fu, Hadi Amata, Xiao Li, Praneeth Chakravarthula, Wolfgang Hei- drich, and Felix Heide. Seeing through obstructions with diffractive cloaking. ACM Transactions on Graphics (TOG), 41(4):1–15, 2022. 3

    2022

  50. [58]

    Differentiable compound optics and processing pipeline op- timization for end-to-end camera design

    Ethan Tseng, Ali Mosleh, Fahim Mannan, Karl St-Arnaud, Avinash Sharma, Yifan Peng, Alexander Braun, Derek Nowrouzezahrai, Jean-Francois Lalonde, and Felix Heide. Differentiable compound optics and processing pipeline op- timization for end-to-end camera design. ACM Transaction...

    2021

  51. [59]

    Extended depth- of-field projector using learned diffractive optics

    Yuqi Li, Qiang Fu, and Wolfgang Heidrich. Extended depth- of-field projector using learned diffractive optics. In 2023 IEEE conference virtual reality and 3D user interfaces (VR), pages 449–459. IEEE, 2023. 2

    2023

  52. [60]

    Deep image deblurring: A survey

    Kaihao Zhang, Wenqi Ren, Wenhan Luo, Wei-Sheng Lai, Bj¨orn Stenger, Ming-Hsuan Yang, and Hongdong Li. Deep image deblurring: A survey. International Journal of Com- puter Vision, pages 1–28, 2022. 2

    2022

  53. [61]

    Cnn-based real-time image restoration in laser sup- pression imaging

    Xiaopeng Peng, Prateek R Srivastava, and Grover A Swartz- lander. Cnn-based real-time image restoration in laser sup- pression imaging. In Optical Sensors and Sensing Congress, pages JTh6A–10. Optica Publishing Group, 2021. 2

    2021

  54. [62]

    Randomized apertures: high reso- lution imaging in far field

    Xiaopeng Peng, Garreth J Ruane, Marco B Quadrelli, and Grover A Swartzlander. Randomized apertures: high reso- lution imaging in far field. Optics express, 25(15):18296– 18313, 2017

    2017

  55. [63]

    Image restoration from a sequence of random masks

    Xiaopeng Peng, Garreth J Ruane, Alexandra B Artusio- Glimpse, and Grover A Swartzlander Jr. Image restoration from a sequence of random masks. In Computational Imag- ing XIII, volume 9401, pages 111–123. SPIE, 2015. 2

    2015

  56. [64]

    Swinir: Image restoration using swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833– 1844, 2021. 2

    2021

  57. [65]

    Uformer: A gen- eral u-shaped transformer for image restoration

    Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A gen- eral u-shaped transformer for image restoration. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683–17693, 2022

    2022

  58. [66]

    Restormer: Efficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728–5739,

  59. [67]

    Pay at- tention to mlps

    Hanxiao Liu, Zihang Dai, David So, and Quoc V Le. Pay at- tention to mlps. Advances in Neural Information Processing Systems, 34:9204–9215, 2021. 2

    2021

  60. [68]

    Maxim: Multi-axis mlp for image processing

    Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5769–5780, 2022. 2

    2022

  61. [69]

    Mamba: Linear-time sequence modeling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2

    2023 arXiv

  62. [70]

    Vision mamba: Efficient visual representation learning with bidirectional state space model

    Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In Forty-first International Conference on Machine Learning. 2

  63. [71]

    U-mamba: Enhancing long-range dependency for biomedical image segmentation

    Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024

    2024 arXiv

  64. [72]

    Neural-driven image editing

    Pengfei Zhou, Jie Xia, Xiaopeng Peng, Wangbo Zhao, Zi- long Ye, Zekai Li, Suorong Yang, Jiadong Pan, Yuanxiang Chen, Ziqiao Wang, et al. Neural-driven image editing. arXiv preprint arXiv:2507.05397, 2025. 2

    2025

  65. [73]

    Spectral representations for convolutional neural networks

    Oren Rippel, Jasper Snoek, and Ryan P Adams. Spectral representations for convolutional neural networks. Advances in neural information processing systems, 28, 2015. 2

    2015

  66. [74]

    Fast fourier convolu- tion

    Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolu- tion. Advances in Neural Information Processing Systems, 33:4479–4488, 2020

    2020

  67. [75]

    Fourier features let networks learn high frequency functions in low dimen- sional domains

    Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ra- mamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimen- sional domains. Advances in Neural Information ...

    2020

  68. [76]

    Deep stacked hierarchical multi-patch network for image deblurring

    Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5978–5986, 2019. 2

    2019

  69. [77]

    Multi-stage progressive image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821–14831, 2021. 2

    2021

  70. [78]

    Removing diffraction image artifacts in under-display camera via dynamic skip connection network

    Ruicheng Feng, Chongyi Li, Huaijin Chen, Shuai Li, Chen Change Loy, and Jinwei Gu. Removing diffraction image artifacts in under-display camera via dynamic skip connection network. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 66...

    2021

  71. [79]

    An intriguing failing of convolutional neural networks and the coordconv solution

    Rosanne Liu, Joel Lehman, Piero Molino, Felipe Pet- roski Such, Eric Frank, Alex Sergeev, and Jason Yosinski. An intriguing failing of convolutional neural networks and the coordconv solution. Advances in neural information processing systems, 31, 2018. 2

    2018

  72. [80]

    Coco- gan: Generation by parts via conditional coordinating

    Chieh Hubert Lin, Chia-Che Chang, Yu-Sheng Chen, Da- Cheng Juan, Wei Wei, and Hwann-Tzong Chen. Coco- gan: Generation by parts via conditional coordinating. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4512–4521, 2019

    2019

  73. [81]

    Infinitygan: Towards infinite-pixel image synthesis

    Chieh Hubert Lin, Hsin-Ying Lee, Yen-Chi Cheng, Sergey Tulyakov, and Ming-Hsuan Yang. Infinitygan: Towards infinite-pixel image synthesis. In International Conference on Learning Representations, 2021. 2

    2021

  74. [82]

    Auto-encoding varia- tional bayes

    Diederik P Kingma and Max Welling. Auto-encoding varia- tional bayes. stat, 1050:1, 2014. 2

    2014

  75. [83]

    Glow: Generative flow with invertible 1x1 convolutions

    Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. Advances in neural information processing systems, 31, 2018. 2

    2018

  76. [84]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Confer- ence on Machine Learning, pages 2256–2265. PMLR, 2015. 2

    2015

  77. [85]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2

    2022

  78. [86]

    Tackling the generative learning trilemma with denoising diffusion gans

    Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. In International Conference on Learning Representa- tions, 2021. 2

    2021

  79. [87]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural informa- tion processing systems, 34:8780–8794, 2021. 2

    2021

  80. [88]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2

    2021

  81. [89]

    Fast sampling of diffusion models via operator learning

    Hongkai Zheng, Weili Nie, Arash Vahdat, Kamyar Azizzade- nesheli, and Anima Anandkumar. Fast sampling of diffusion models via operator learning. In International Conference on Machine Learning, pages 42390–42402. PMLR, 2023. 2

    2023

  82. [90]

    On distillation of guided diffusion models

    Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297–14306, 2023. 2

    2023

  83. [91]

    Mst++: Multi-stage spectral-wise transformer for efficient spectral reconstruction

    Yuanhao Cai, Jing Lin, Zudi Lin, Haoqian Wang, Yulun Zhang, Hanspeter Pfister, Radu Timofte, and Luc Van Gool. Mst++: Multi-stage spectral-wise transformer for efficient spectral reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogniti...

    2022

  84. [92]

    com / rafael - fuente / diffractsim

    https : / / github . com / rafael - fuente / diffractsim. 4

  85. [93]

    Physically-based real-time lens flare rendering

    Matthias Hullin, Elmar Eisemann, Hans-Peter Seidel, and Sungkil Lee. Physically-based real-time lens flare rendering. In ACM SIGGRAPH 2011 papers, pages 1–10. 2011. 4

    2011

  86. [94]

    How to train neural networks for flare removal

    Yicheng Wu, Qiurui He, Tianfan Xue, Rahul Garg, Jiawen Chen, Ashok Veeraraghavan, and Jonathan T Barron. How to train neural networks for flare removal. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2239–2247, 2021

    2021

  87. [95]

    Flare7k: A phenomenological night- time flare removal dataset

    Yuekun Dai, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Flare7k: A phenomenological night- time flare removal dataset. Advances in Neural Information Processing Systems, 35:3926–3937, 2022. 4

    2022

  88. [96]

    Adaptive osculatory rational in- terpolation for image processing

    Min Hu and Jieqing Tan. Adaptive osculatory rational in- terpolation for image processing. Journal of Computational and Applied Mathematics, 195(1-2):46–53, 2006. 6, 7

    2006

  89. [97]

    Gradient surgery for multi-task learning

    Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in neural information pro- cessing systems, 33:5824–5836, 2020. 6

    2020

  90. [98]

    Opening: A comprehensive benchmark for judging open-ended interleaved image-text generation

    Pengfei Zhou, Xiaopeng Peng, Jiajun Song, Chuanhao Li, Zhaopan Xu, Yue Yang, Ziyao Guo, Hao Zhang, Yuqi Lin, Yefei He, et al. Opening: A comprehensive benchmark for judging open-ended interleaved image-text generation. In Proceedings of the Computer Vision and Pattern Recognit...

    2025

  91. [99]

    Mdk12-bench: A comprehensive eval- uation of multimodal large language models on multidis- ciplinary exams

    Pengfei Zhou, Xiaopeng Peng, Fanrui Zhang, Zhaopan Xu, Jiaxin Ai, Yansheng Qiu, Chuanhao Li, Zhen Li, Ming Li, Yukang Feng, et al. Mdk12-bench: A comprehensive eval- uation of multimodal large language models on multidis- ciplinary exams. arXiv preprint arXiv:2508.06851, 2025. 7

    2025 arXiv

  92. [100]

    https://unsplash.com/license. 7

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.